We would like you to take the following benchmark scores with a grain of salt, sugar or whatever it is that helps you digest things more easily. Please keep in perspective that the following tests are based on a benchmark suite that performs real-world functions like JPEG decompression, etc. However, the tests are not based on any end user retail applications but rather a "synthetic" collection of specific computational functions.
| PCMark 2002 Testing
| Synthetic benchmarks that show huge potential for Hyperthreading
Here we see that the Pentium 4 3.06GHz scales accordingly in performance versus the 2.8GHz chip, with roughly an 8% lead. However, overall the CPU test does not show any benefit for the Hyperthreading enabled system. That is until we look more closely at the "Crunch" test section.
We asked the folks at MadOnion to give us a little background on what was behind the "Crunch" test in PCMark 2002. Here's what Patric Ojala, Senior Manager, Benchmark Development at MadOnion.com told us.
The Crunch test runs 3 tests simultaneously:
CPU Test - JPEG Decompression, stressing mostly the CPU and somewhat
also the memory subsystem.
Memory Test - Raw 3072 kB Block Modify, stressing mostly the memory
subsystem, but also somewhat the CPU.
Memory Test - Video - 1 scanline, stressing mostly the graphics card
memory bandwidth and the AGP bus.
These tests are all described above, and can be run separately in PCMark2002. The idea of the Crunch test is to see how well the system can perform several tasks simultaneously and stress different parts of the system concurrently. The Crunch test was designed to be a stress test for total system throughput; CPU, Memory and Graphics speed. This kind of measurement should be good for total system comparison, multiprocessor benchmarking, and motherboard chipset efficiency testing. The included tests were selected because:
CPU Test - JPEG Decompression, since it is a CPU intensive test that does a task common for many applications (web, office, image composer). The data set is larger than what fits into the cache, so there is some memory transfer involved, which is appropriate for any CPU test.
Memory Test - Raw 3072 kB Block Modify, where the large block is used to avoid only cache measurement. The block modify functionality is typical for data modification, like a light Photoshop filter. The data operation is light enough to stress above all the memory read and write speed. Still, just data transfer back and forth might open up a chance for caching. Therefore the small modify operation was included.
The Video Memory test above all stresses the AGP bus and the graphics card memory speed. The selected single scanline transfer weights more on the video memory speed than the AGP speed (compared to the multiple scanline tests), but we think this kind of smaller video memory data transfer is more common in PC home and office usage.
The three threads all have normal priority, so we don't try to control which part of the system runs most efficiently. We use pre-set work amounts on all tests, so on many single processor systems you can see how the CPU and memory threads complete their tasks first. The video memory test first scrolls really slowly, while the other tests are running. Then it suddenly picks up speed and the scrolling gets smooth, when only the video memory test remains. No "internally SMP" processors were around when we made PCMark2002.
In every case during the Crunch test, we see the Hyperthreaded Pentium 4 blow by it's non HT enabled counterparts with ease. The Crunch Blit test shows orders of magnitude more performance for the Pentium 4 3.06GHz with HT enabled. This could indeed be a foreshadow of things to come with respect to multithreaded scenarios, like those that were arranged in the PCMark 2002 Crunch test. Remember, the folks at MadOnion did not specifically code the benchmark to take advantage of the HT enabled P4. They told us the chip wasn't even around when they designed their benchmark. Regardless, synthetic benchmarks need to be kept in perspective as the relative performance metrics that they are.
Let's get back to the "real world". Well sort of...
Quake 3 Time Demos and UT